testsAndMisc/docs/shell-split-verification.md

Shell split verification: how a green split is still broken

Split out of docs/shell-split-recipes.md when that file hit the same 250-line cap it exists to serve. That file covers how to make a split; this one covers how to find out it is wrong. Read both before the next linux_configuration or phone_focus_mode split.

Every section below cost a failed run, a revert, or — once — a live system mutation.

A --skip-install flag is not evidence the script is safe to run

shell_check.sh gates only install_if_missing's pacman branch on SKIP_INSTALL. The AUR branch inside install_linters runs regardless, so a backgrounded "baseline" run installed four packages before it was killed mid-makepkg. Decision 6 was violated by a flag that looked like permission.

Before running any script to capture a baseline, grep the flag's variable and confirm it gates every mutation path, not just the first one:

grep -n 'SKIP_INSTALL\|DRY_RUN' <script>   # then read each hit

--list-only was the genuinely safe path here, because it returns before install_linters is reached at all.

Check whether the target is DEPLOYED as a single file — the trace cannot

Before designing seams, find out whether some other script copies the target somewhere and runs it from there. An entry+lib/ shape ships an entry whose deployed SCRIPT_DIR has no lib/, so it dies on its first source under set -e — while your trace stays green, because the trace runs the repo copy where lib/ exists.

grep -rn '<target-basename>' linux_configuration/ --include='*.sh' \
  | grep -v '<the target itself>:'

Confirmed blocked on this, not theoretically: install_leechblock.sh (the pacman installer cps it to /usr/local/share/digital_wellbeing/ and pacman_wrapper.sh:831 prefers that copy on every pacman invocation) and block_compulsive_opening.sh (install_all copies itself to /usr/local/bin). A target invoked as $CONFIG_DIR/scripts/... is fine.

A stubbed run cannot reach a bug in a seam that passes state

meta/scripts/verify_shell_split.sh and the hand-rolled declare -F probe both source the libs and never call the functions. That is enough to prove every source line resolves, and nothing more. The analyze_repo.sh split passed bash -n, shellcheck, a function-set diff, a line-set diff and the stubbed probe — and still shipped a script that aborted immediately after language detection. Two separate bugs, neither reachable without executing:

• set -e and the function tail. Eight ((LANG_FILES[x] > 0)) && HAS_Y=true lines were top-level statements, where a false && merely leaves $? non-zero. Wrapped into a function they become its return value, and the last one (java, false for most repos) killed the script. Any time you wrap former top-level code in a function, check what its last statement is; append || true if it is a bare conditional.
• A self-referencing nameref. local -n LANG_CODE_FILES="$1" pointing at a global of the same name still moves data on bash 5.3, but warns circular name reference on every access — straight into the script's output. Name the local something else (_code_files).

So: when a seam passes state, run the thing. Stub only the parts that mutate the system, and diff the generated artifacts against a detached worktree at the pre-split commit. That comparison is what proved the fix — same file set, every text artifact byte-identical.

The harness and the fixtures

Everything about running the check — trace_shell_split.sh, the --prefix mode for scripts whose job is placing files, the regression fixtures, the entry-point and shfmt traps — moved to docs/shell-split-harness.md when this file hit the same 250-line cap it exists to serve. This file stays the why; that one is the how.

Prove the move was verbatim: meta/scripts/verify_shell_split.sh

The Python splits in this repo are verified with a throwaway ast identity check. There is no ast for bash, so the shell equivalent normalises through shfmt -mn (minify: strips comments and indentation, keeps logic) and compares a hash per top-level function:

bash meta/scripts/verify_shell_split.sh HEAD <old-path> <new-path>...
# IDENTICAL: all N top-level functions moved verbatim

It reports which function changed, so a diff points at the seam rather than the whole file. Run it after every split, and again after shfmt/pre-commit autofixes — reordering an import or reflowing a line is exactly what it is designed to see through.

Verify the harness before trusting it. A normalizer that silently passes everything is worse than none. The three checks it must survive:

input	expected
unchanged copy of the file	IDENTICAL
comments/blank lines/indentation changed	IDENTICAL
one character of logic changed ("$1" → "$2")	DIFFERENCE FOUND, naming the function

The first attempt at that mutation test used sed 's/exit 1/exit 2/' on a file containing no exit 1. It reported IDENTICAL — not because the harness was broken, but because the mutation never applied. Confirm the mutation actually changed the file (diff it) before concluding anything about the harness.

Deployment triage of the over-cap scripts (2026-08-17)

Run before designing any seam, per the section above. deploy.sh copies a hardcoded per-file list into /data/local/tmp/focus_mode and never deploys lib/:

• phone_focus_mode/* (8 over-cap scripts) — deployed individually. A new sibling must be added to deploy.sh's copy list in the same commit, or focus mode dies on the phone at its first source. These scripts source only config.sh, which is on that list, so a sibling-file seam is workable — an entry+lib/ seam is not.
• install_leechblock.sh, block_compulsive_opening.sh — copied to /usr/local/… and preferred there by pacman_wrapper.sh:831 on every pacman invocation. Teach the installer to deploy the directory first, in its own commit. Highest blast radius; do these last.
• hosts/install.sh, setup_hosts_guard.sh, lib/monitor.sh — install helper scripts to /usr/local/sbin or /usr/local/bin; check whether the file being split is itself the deployed one.
• The other 16 (fresh-install/main.sh, nvidia_troubleshoot.sh, setup_passwordless_system.sh, clean_audio.sh, enforce_vbox_hosts.sh, migrate_hosts_guard_to_guard_lib.sh, libre_translate.sh, diagnose_pacman_hook_stall.sh, install_plagiarism_tools.sh, steam_compatibility.sh, deploy.sh, setup_night_lockdown.sh, pacman_wrapper.sh, generate_study_materials.sh, check_and_enable_services.sh, setup_midnight_shutdown.sh) — no deployed copy found, so an entry+lib/ shape is safe.

A lib must contain the readers of every global it assigns

install_pacman_wrapper.sh:29 records the constraint: a definitions-only lib "assigns without referencing, which is SC2034, and the repo forbids suppressions." The rule that follows from it is not "assignments stay in the entry script" — it is that a file must not assign a variable it never reads. A lib holding both the writer and its readers is clean.

Getting this inverted is easy and costs several revert cycles. libre_translate.sh was split with parse_args (the writer) left in the entry and health_check / sample_request / start_container_ephemeral (the readers) moved to the lib — the worst of both, and SC2034 fired on nine globals. The fix is to colocate: move the writer and its consumers into the same file.

The pre-commit hook is shellcheck with no -x, so each file is analysed standalone and a # shellcheck source= directive will not make the checker follow the link. Verify by running shellcheck <lib> on its own before committing; the hook will not be any more forgiving.

This is the shell analogue of the monkeypatch.setattr target trap: the question is always which file's namespace owns the name.

libre_translate.sh (488) is parked. What remains after its clean seams are taken is parse_args (111 lines), write_env_file, detect_container_user and main, which between them assign and read the same configuration globals. Getting the entry under the cap needs parse_args restructured rather than relocated — a refactor, not a verbatim move, so it is out of scope for this method. Attempt it with a test harness, or accept an entry that stays over.